Нейросеть планирует и воображает как мозг: работа в Nature Machine Intelligence

A Neural Network Plans and Imagines Like a Brain: Research in Nature Machine Intelligence

Researchers led by Lin and colleagues published a generative model in the journal Nature Machine Intelligence on July 21, 2026, inspired by how the brain works: it plans actions and solves tasks by relying on cognitive maps and learning solely through local synaptic plasticity.

What the model is built on

The model by Lin and colleagues combines three mechanisms, each borrowed from neuroscience. At its core are cognitive maps — internal representations of the environment by which the system "charts a route" to a goal even before taking real action.

• Publication — July 21, 2026, journal Nature Machine Intelligence (DOI 10.1038/s42256-026-01254-4)
• Authors — Lin and colleagues (Lin et al.)
• Three pillars of the model: cognitive maps, stochastic computation, compositional coding
• Learning — only local synaptic plasticity, without global backpropagation
• Two functions of intelligence: planning and problem solving

Stochastic computation means the model operates through random sampling (neural sampling): it doesn't produce one rigid answer, but runs through possible versions of the future. Compositional coding allows it to assemble new scenes and plans from familiar elements — just as a person imagines a situation they have never been in.

How this differs from ordinary neural networks

The model by Lin and colleagues differs from most neural networks in its method of learning. Modern systems learn via backpropagation, which requires global access to all the network's weights and is biologically implausible. Here, each connection instead updates according to a local rule — based on signals from its nearest neighbors, as presumably happens in the living brain.

Local learning makes the architecture both closer to biology and potentially more economical for neuromorphic hardware, where global operations are costly. According to the authors, the model delivers two key traits of intelligence at once — something ordinary generative systems only imitate partially.

"The model provides two key features of intelligence: planning and problem solving," — from the article by

Lin et al. in Nature Machine Intelligence.

What goal-directed imagination is

Goal-directed imagination is the model's ability to mentally "play out" paths to a given goal. Rather than simply reacting to the current input, the system samples possible sequences of actions from the cognitive map and selects those that lead to the goal.

This approach brings planning in machines closer to how the brain operates: first an imagined route, then a real step. It is precisely the combination of three elements — a cognitive map, stochastic sampling, and local learning — that sets this work apart from purely engineered planners, which make no claim to biological plausibility and typically rely on global learning.

What this means

The work by Lin and colleagues is a step toward neural networks that plan and imagine according to the principles of the brain, rather than merely recognizing patterns. The combination of local learning and stochastic inference fits well with energy-efficient neuromorphic chips and robotics, where the two stated functions — planning and problem solving — are especially prized.